AI Insights: Approvals Still Pending – Learning From Our AI Mistakes

Let me tell you a little story about AI.

A couple of years back, Mindera was involved in a project. It was outcome-based and had to be delivered in a short time frame. So, we thought: “AI is booming, let’s deliver it through AI.”

In theory this would shrink timelines, accelerate processes and we’d have an awesome case study! The reality proved quite the opposite.

In fact, it could have been a disaster. But with persistence and determination, learning as we went along, we got through it. It’s certainly not an experience we want to repeat, but we learned a huge amount about applying AI. Here’s what happened, so you can avoid the same mistakes.

Despite adopting AI tools like Microsoft Copilot and AI Foundry, our team experienced a 27% schedule overrun, extending from five months to eight months and requiring an increase in team size from five to seven engineers.

So where did we go wrong?

The key causes of friction

The project was derailed not by AI itself, but by four related factors that we failed to anticipate due to our lack of experience:

Organisational naivety: Since this was not purely an AI initiative, we could not influence how other teams, that we were dependent on, worked. So, their speed to carry out reviews directly affected our progress. We wrongly assumed that AI could be adopted through simple workshops without changing existing ways of working or accounting for necessary dependencies.

Bottlenecks: Things started well. We began writing code, pushing it into several repositories. The one we owned was progressing quickly. Unfortunately, the others were not. This cost us a 27% overrun. The primary bottlenecks were not the AI-driven code generation, but rather the approval processes, dependency management across multiple repositories, and the time required to onboard and configure AI models.

External dependencies: The project relied on teams with priorities that were not aligned with those of our project. This resulted in delays for code reviews and necessary approvals. We had pull requests waiting for days, because we needed other teams to review what we’d done, adding to their workload, yet this was not their priority. The reviews were only minor, but still vital and the other teams could choose if and when they carried them out.

Governance and legal hurdles: Our team faced significant delays due to late-stage legal reviews regarding the use of global versus regional AI models. Essentially, we were paying the price for being the first to use the models in the organisation. Something we should have considered. Furthermore, the AI Foundry model we were using was decommissioned without warning. Our product was in the testing phase at the time, about to be released, and it suddenly stopped working. This was a failure because there was no governance in place to warn us in advance of this critical incident.

Lessons learned

Despite the barrage of problems we encountered, we persisted and learned from them, finally getting the outcome we originally set out to achieve. It just took a little longer than expected. Yes, we all had a few more grey hairs! But we certainly gained from the experience. The two big lessons were that you can’t deploy AI successfully without:

Coaching and Champions: Our team improved by shifting from one-off workshops to ongoing coaching and appointing an ‘AI champion’ to lead by example and demonstrate how to use AI tools in the most effective way.

Alignment: Throughput significantly increased after our team invested weeks in aligning with other departments regarding ways of working and establishing a collaborative approach to handling dependencies and reviews.

Getting there… eventually

What we achieved in a somewhat roundabout way does show the power of AI to transform projects when applied in the right way. By the end of the process, our team achieved a 0.7-hour average review time, released the project with zero bugs in production, and increased output from 3 to 52 pull requests merged per week.

The overall insight gained from this experience was that AI is an ‘amplifier’ rather than a standalone solution. It only works if the surrounding organisational ecosystem is properly aligned. While the technology is important, AI success hinges on coaching, rigorous dependency mapping, and treating governance as a ‘first-class citizen’ within the workflows. Something that we’ve taken forward and has proved key to driving success in our AI projects ever since.

Dive deeper into the learnings of this AI project by watching Pedro Vicente's full Mindera AI Insights 2026 presentation below.